Orpheus-TTS local speech synthesizer written entirely in C#

written from : isaiahbjork/orpheus-tts-local

Requirements

.NET 9
Cuda 12

Usage

1. Clone the git repo:

• git clone https://github.com/TheAjaykrishnanR/TaraSharp

2. Download the the quantized model from here

3. Download the decoder weights from here and place it in snac\24khz alongside config.json. If you want to use the original weights, see conversion method below.

4. cd TaraSharp

5. Change model_path in Program.cs to the location where your gguf file is.

6. dotnet run

Additional Info

The original snac model available here was created using PyTorch. For the most part models created in PyTorch can be loaded without much hassle in TorchSharp (not vice versa). However the snac decoder uses the weighted_norm() function to wrap Conv1d layers which is not implemented in TorchSharp as of today. weighted_norm() decouples weights into their magnitude and direction. So every Conv1d layer's weight is now split into two : parametrization.weight.original0 and parametrization.weight.original1. So when this model is to be loaded into TorchSharp the weights have to be reconstructed in the state_dict and the above two keys deleted. scripts\make_unnormed_torchsharp.py does precisely that.